sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 226 insertions, 0 deletions

diff --git a/net/rds/tcp_send.c b/net/rds/tcp_send.c
new file mode 100644
index 000000000..7c52acc74
--- /dev/null
+++ b/net/rds/tcp_send.c
@@ -0,0 +1,226 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <net/tcp.h>
+
+#include "rds_single_path.h"
+#include "rds.h"
+#include "tcp.h"
+
+void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	tcp_sock_set_cork(tc->t_sock->sk, true);
+}
+
+void rds_tcp_xmit_path_complete(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	tcp_sock_set_cork(tc->t_sock->sk, false);
+}
+
+/* the core send_sem serializes this with other xmit and shutdown */
+static int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
+{
+	struct kvec vec = {
+		.iov_base = data,
+		.iov_len = len,
+	};
+	struct msghdr msg = {
+		.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL,
+	};
+
+	return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len);
+}
+
+/* the core send_sem serializes this with other xmit and shutdown */
+int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
+		 unsigned int hdr_off, unsigned int sg, unsigned int off)
+{
+	struct rds_conn_path *cp = rm->m_inc.i_conn_path;
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct msghdr msg = {};
+	struct bio_vec bvec;
+	int done = 0;
+	int ret = 0;
+
+	if (hdr_off == 0) {
+		/*
+		 * m_ack_seq is set to the sequence number of the last byte of
+		 * header and data.  see rds_tcp_is_acked().
+		 */
+		tc->t_last_sent_nxt = rds_tcp_write_seq(tc);
+		rm->m_ack_seq = tc->t_last_sent_nxt +
+				sizeof(struct rds_header) +
+				be32_to_cpu(rm->m_inc.i_hdr.h_len) - 1;
+		smp_mb__before_atomic();
+		set_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags);
+		tc->t_last_expected_una = rm->m_ack_seq + 1;
+
+		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
+			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
+
+		rdsdebug("rm %p tcp nxt %u ack_seq %llu\n",
+			 rm, rds_tcp_write_seq(tc),
+			 (unsigned long long)rm->m_ack_seq);
+	}
+
+	if (hdr_off < sizeof(struct rds_header)) {
+		/* see rds_tcp_write_space() */
+		set_bit(SOCK_NOSPACE, &tc->t_sock->sk->sk_socket->flags);
+
+		ret = rds_tcp_sendmsg(tc->t_sock,
+				      (void *)&rm->m_inc.i_hdr + hdr_off,
+				      sizeof(rm->m_inc.i_hdr) - hdr_off);
+		if (ret < 0)
+			goto out;
+		done += ret;
+		if (hdr_off + done != sizeof(struct rds_header))
+			goto out;
+	}
+
+	while (sg < rm->data.op_nents) {
+		msg.msg_flags = MSG_SPLICE_PAGES | MSG_DONTWAIT | MSG_NOSIGNAL;
+		if (sg + 1 < rm->data.op_nents)
+			msg.msg_flags |= MSG_MORE;
+
+		bvec_set_page(&bvec, sg_page(&rm->data.op_sg[sg]),
+			      rm->data.op_sg[sg].length - off,
+			      rm->data.op_sg[sg].offset + off);
+		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1,
+			      rm->data.op_sg[sg].length - off);
+		ret = sock_sendmsg(tc->t_sock, &msg);
+		rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
+			 rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
+			 ret);
+		if (ret <= 0)
+			break;
+
+		off += ret;
+		done += ret;
+		if (off == rm->data.op_sg[sg].length) {
+			off = 0;
+			sg++;
+		}
+	}
+
+out:
+	if (ret <= 0) {
+		/* write_space will hit after EAGAIN, all else fatal */
+		if (ret == -EAGAIN) {
+			rds_tcp_stats_inc(s_tcp_sndbuf_full);
+			ret = 0;
+		} else {
+			/* No need to disconnect/reconnect if path_drop
+			 * has already been triggered, because, e.g., of
+			 * an incoming RST.
+			 */
+			if (rds_conn_path_up(cp)) {
+				pr_warn("RDS/tcp: send to %pI6c on cp [%d]"
+					"returned %d, "
+					"disconnecting and reconnecting\n",
+					&conn->c_faddr, cp->cp_index, ret);
+				rds_conn_path_drop(cp, false);
+			}
+		}
+	}
+	if (done == 0)
+		done = ret;
+	return done;
+}
+
+/*
+ * rm->m_ack_seq is set to the tcp sequence number that corresponds to the
+ * last byte of the message, including the header.  This means that the
+ * entire message has been received if rm->m_ack_seq is "before" the next
+ * unacked byte of the TCP sequence space.  We have to do very careful
+ * wrapping 32bit comparisons here.
+ */
+int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack)
+{
+	if (!test_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags))
+		return 0;
+	return (__s32)((u32)rm->m_ack_seq - (u32)ack) < 0;
+}
+
+void rds_tcp_write_space(struct sock *sk)
+{
+	void (*write_space)(struct sock *sk);
+	struct rds_conn_path *cp;
+	struct rds_tcp_connection *tc;
+
+	read_lock_bh(&sk->sk_callback_lock);
+	cp = sk->sk_user_data;
+	if (!cp) {
+		write_space = sk->sk_write_space;
+		goto out;
+	}
+
+	tc = cp->cp_transport_data;
+	rdsdebug("write_space for tc %p\n", tc);
+	write_space = tc->t_orig_write_space;
+	rds_tcp_stats_inc(s_tcp_write_space_calls);
+
+	rdsdebug("tcp una %u\n", rds_tcp_snd_una(tc));
+	tc->t_last_seen_una = rds_tcp_snd_una(tc);
+	rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc), rds_tcp_is_acked);
+
+	rcu_read_lock();
+	if ((refcount_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf &&
+	    !rds_destroy_pending(cp->cp_conn))
+		queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 0);
+	rcu_read_unlock();
+
+out:
+	read_unlock_bh(&sk->sk_callback_lock);
+
+	/*
+	 * write_space is only called when data leaves tcp's send queue if
+	 * SOCK_NOSPACE is set.  We set SOCK_NOSPACE every time we put
+	 * data in tcp's send queue because we use write_space to parse the
+	 * sequence numbers and notice that rds messages have been fully
+	 * received.
+	 *
+	 * tcp's write_space clears SOCK_NOSPACE if the send queue has more
+	 * than a certain amount of space. So we need to set it again *after*
+	 * we call tcp's write_space or else we might only get called on the
+	 * first of a series of incoming tcp acks.
+	 */
+	write_space(sk);
+
+	if (sk->sk_socket)
+		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
+}